## Summary of mortality-temperature regression models

This section summarizes the mortality-temperature regression models in Carleton et al. (2022), which are estimated under five robustness specifications with the following assumptions regarding fixed effects, regression method, and data construction:

|                           |               Spec. 1               |                Spec. 2                 |                Spec. 3                 |                Spec. 4                 |                Spec. 5                 |
| :------------------------ | :---------------------------------: | :------------------------------------: | :------------------------------------: | :------------------------------------: | :------------------------------------: |
| Functional form           |        4th order polynomial         |          4th order polynomial          |          4th order polynomial          |          4th order polynomial          |          4th order polynomial          |
| Fixed effects             | (Age x ADM2)  & (Country x Year) FE | (Age x ADM2)  & (Age x Country x Year) | (Age x ADM2)  & (Age x Country x Year) | (Age x ADM2)  & (Age x Country x Year) | (Age x ADM2)  & (Age x Country x Year) |
| Time trend                |                  -                  |                   -                    |        Age x ADM1 linear trend         |                   -                    |                   -                    |
| Regression method         |                 OLS                 |                  OLS                   |                  OLS                   |                  FGLS                  |                  OLS                   |
| Weather data construction |          12-month exposure          |           12-month exposure            |           12-month exposure            |           12-month exposure            |           13-month exposure            |

Note that Specification 2 is the "preferred" model in the analysis. ADM0, ADM1, and ADM2 correspond roughly to country, state, and county administrative units.

The following list provides details on the models in the order in which they appear in the paper. Note that additional detail on the specific inputs and outputs of each script can be found in script headers.

#### 1. All-age mortality-temperature response function estimated using pooled subnational data.

- Location in Carleton et al. (2022): Table D2

* Folder: `1_age_combined/`
  * `age_combined_regressions.do` - estimates the model.
  * `age_combined_displayresults.do` - generates Table D2.

- Description: 
  - 4th order polynomial in daily average temperature (GMFD) for the five specifications summarized above.
  - All specs are population weighted regressions with ADM1 unit standard error clustering and AGE x ADM0 precipitation controls.

#### 2. Mortality-temperature response functions for <5, 5-64, and >64 age groups estimated using pooled subnational data.

- Location in Carleton et al. (2022): Table D3

* Folder: `2_age_spec/`
    * `age_spec_regressions.do` - estimates the model.
    * `age_spec_displayresults.do` - generates Table D3.

- Description: 
    - 4th order polynomial in daily average temperature (GMFD) for the five specifications summarized above.
    - All specs are population weighted regressions with ADM1 unit standard error clustering and AGE x ADM0 precipitation controls.

#### 3. Mortality-temperature response functions for <5, 5-64, and >64 age groups accounting for spatial heterogeneity in average income and climate.

- Locations in Carleton et al. (2022): 
    - Figures I, D1, D2 and Table D3
    - In-text discussion of the reduction in temperature sensitivity associated with moving between terciles of the interaction space.
    - All projection output is based upon the interaction model estimated here.

* Folder: `3_age_spec_interacted/`
    * `age_spec_interacted_regressions.do` - estimates the model.
    * `age_spec_interacted_array_plots_presentation.do`- generates array plots in Figures I, D1, and D2.
    * `age_spec_interacted_displayresults.do` - generates Table D3.
    * `array_output_in-text.do` - generates in-text summary statistics on the changes in temperature-sensitivity associated with changes in income and long-run average temperature.
    * `age_spec_interacted_genCSVV.do` - generates configuration file (or "CSVV") of the main specification for the Climate Impact Lab projection system.
        * See the next step, `2_projection` for a more detailed discussion of this input and how it's implemented in the projection system.

#### 4. Alternative regression models

- Locations in Carleton et al. (2022): 
    - Tables D3 and D4, and Figures D4, D6, D7, D8

* Folder: `4_alternative_models/`
    * `1_regressions/` - scripts to estimate alternative regression models
    * `2_analysis/`- scripts to plot figures and/or tables displaying regression results
    * `institutional_covariates/` - scripts to estimate and plot the alternative models that include additional determinants of heterogeneity, such as institutional quality and workforce informality

#### 5. Burgess et al. (2017) out-of-sample exercise

- Locations in Carleton et al. (2022):
    - Table D6 and Figure D11

* Folder: `5_Burgess_India_test/`
    * `age_combined_india-response_regressions.do` computes the "true" Indian all-age response function, as closely replicating Burgess et al. (2017) as possible (see notes on replicability at the top of the script)
    * `age_spec_interacted_india_compare_responses.do` uses three versions of the main interaction model in Carleton et al. (2022) to predict the mortality-temperature response across India, for comparison with the Burgess et al. result
    * `compute_average_age_share_IND.R` and `format_IND_tavg_bins.R` are data formatting scripts
    * `Table_D6_india_compare_table.R` and `Figure_D11_India_response_compare.do` plot results in Table D6 and Figure D11, respectively

